JNR Deforestation Risk Maps

Abstract

The following starter script documents two workflows for deriving deforestation risk maps in accordance with Verra’s new methodology for unplanned deforestation allocation in jurisdictional nested REDD+ projects using the VT0007 toolset.

Keywords

REDD+, VCS, Verra, Carbon verification, Jurisdictional

Summary

The following details two workflow approaches to Verra’s recommended sequence of deforestation risk map development Verra (2021). For comparison purposes, both workflows use the same training sample (Stanimirova et al. 2023) and STAC-formatted Landsat Collection-2 Level-2 imagery.

Workflow-1 is coded in R and is recommended for smaller areas of analysis, as it offers additional functions for model tuning and classifer evaluation. Workflow-2, which is coded in Python and Google Earth Engine functions, is recommended for larger areas of interest (Update java link here).

Figure 1: Verra’s recommended risk map development sequence (VT0007:6)

1. Workflow in R -> sits

1.1 Testing phase

import numpy as np
import sys 
print(sys.executable)

/home/seamus/Repos/VT0007-deforestation-map/bin/python

Jurisdictional boundaries

# assign master crs
crs_master  = sf::st_crs("epsg:4326")
aoi_country = geodata::gadm(country="GUY", level=0, path=tempdir()) |>
  sf::st_as_sf() |> sf::st_cast() |> sf::st_transform(crs_master)
aoi_states = geodata::gadm(country="GUY", level=1, path=tempdir()) |>
  sf::st_as_sf() |> sf::st_cast() |> sf::st_transform(crs_master) |>
  dplyr::rename(State = NAME_1)
aoi_target = dplyr::filter(aoi_states, State == "Barima-Waini")

tmap::tmap_mode("view")
tmap::tm_shape(aoi_states) + tmap::tm_borders(col = "white", lwd = 0.5) +
  tmap::tm_text("State", col = "white", size = 1, alpha = 0.3, just = "bottom") +
  tmap::tm_shape(aoi_country) + tmap::tm_borders(col = "white", lwd = 1) +
  tmap::tm_shape(aoi_target) + tmap::tm_borders(col = "red", lwd = 2) +
  tmap::tm_text("State", col = "red", size = 1.3) +
  tmap::tm_basemap("Esri.WorldImagery")

Assemble HRP time series

We assemble and process a data cube representing a historical reference period (HRP) between 2013-01-01 and 2023-01-01 for the country of Suriname. Using the sits_regularize functions, we apply a cloud masking and pixel back-filling based on cloudless ranking and median-normalization across 5-year intervals to derive three dry-season mosaics for 2013, 2018 and 2023.

# 2014 -------------------
# cloud-assemble data cube
cube_raw_2014 = sits::sits_cube(
  source      = "MPC",
  collection  = "LANDSAT-C2-L2",
  bands       = c("RED", "GREEN", "BLUE", "NIR08", "SWIR16", "CLOUD"),
  roi         = aoi_target,
  start_date  = as.Date("2014-01-01"),
  end_date    = as.Date("2014-07-01"),
  progress    = T
  )

# regularize data cube
cube_reg_2014 = sits::sits_regularize(
  cube        = cube_raw_2014,
  roi         = aoi_target,
  res         = 60,
  period      = "P180D",
  output_dir  = here::here("cubes", "reg", "2014"),
  memsize     = 16,
  multicores  = 8,
  progress    = T
  )

# 2019 -------------------
# cloud-assemble data cube
cube_raw_2019 = sits::sits_cube(
  source      = "MPC",
  collection  = "LANDSAT-C2-L2",
  bands       = c("RED", "GREEN", "BLUE", "NIR08", "SWIR16", "CLOUD"),
  roi         = aoi_target,
  start_date  = as.Date("2019-01-01"),
  end_date    = as.Date("2019-07-01"),
  progress    = T
  )

# regularize data cube
cube_reg_2019 = sits::sits_regularize(
  cube        = cube_raw_2019,
  roi         = aoi_target,
  res         = 60,
  period      = "P180D",
  output_dir  = here::here("cubes", "reg", "2019"),
  memsize     = 16,
  multicores  = 8,
  progress    = T
  )

# 2024 -------------------
# cloud-assemble data cube
cube_raw_2024 = sits::sits_cube(
  source      = "MPC",
  collection  = "LANDSAT-C2-L2",
  bands       = c("RED", "GREEN", "BLUE", "NIR08", "SWIR16", "CLOUD"),
  roi         = aoi_target,
  start_date  = as.Date("2024-01-01"),
  end_date    = as.Date("2024-07-01"),
  progress    = T
  )

# regularize data cube
cube_reg_2024 = sits::sits_regularize(
  cube        = cube_raw_2024,
  roi         = aoi_target,
  res         = 60,
  period      = "P180D",
  output_dir  = here::here("cubes", "reg", "2024"),
  memsize     = 16,
  multicores  = 8,
  progress    = T
  )

# plot cube timelines
sits_timeline(cube_reg_2014)
sits_timeline(cube_reg_2019)
sits_timeline(cube_reg_2024)
plot(cube_reg_2014,
  red         = "RED",
  green       = "GREEN",
  blue        = "BLUE",
  date        = "2014-01-03"
  )

plot(cube_reg_2019,
  red         = "RED",
  green       = "GREEN",
  blue        = "BLUE",
  date        = "2019-01-08"
  )

plot(cube_reg_2024,
  red         = "RED",
  green       = "GREEN",
  blue        = "BLUE",
  date        = "2024-01-07"
  )

Classify HRP time series

We import the GLanCE training dataset of annual times series points that includes 7 land cover classes (Figure 2; (Woodcock et al., n.d.)). Training samples are fitted to a Random Forest model and post-processed with a Bayesian smoothing and then evaluated using confusion matrix. The classifier is then calibrated by mapping pixel uncertainty, adding new samples in areas of high uncertainty, reclassifying with improved samples and re-evaluated using confusion matrix.

Figure 2: Land cover classes included in the GLanCE Level 1 classification scheme (Woodcock et al 2022)

# Extract training set from gee to drive & import: https://gee-community-catalog.org/projects/glance_training/?h=training 
glance_training_url = "https://drive.google.com/file/d/1CgBP2J2OdOhmOiVS4hGibLEMyVLTe1_P/view?usp=drive_link"
# file_name = "glance_training.csv"
# download.file(url = url, path = here::here("training"), destfile = file_name)
glance_training = read.csv(here::here("training", "glance_training.csv"))

data("samples_prodes_4classes")
# Select the same three bands used in the data cube
samples_4classes_3bands <- sits_select(
  data = samples_prodes_4classes,
  bands = c("B02", "B8A", "B11")
  )

# Train a random forest model
rfor_model <- sits_train(
  samples = samples_4classes_3bands,
  ml_method = sits_rfor()
  )

# Classify the small area cube
s2_cube_probs <- sits_classify(
  data = s2_reg_cube_ro,
  ml_model = rfor_model,
  output_dir = "./cubes/02_class/",
  memsize = 15,
  multicores = 5
  )

# Post-process the probability cube
s2_cube_bayes <- sits_smooth(
  cube = s2_cube_probs,
  output_dir = "./cubes/02_class/",
  memsize = 16,
  multicores = 4
  )

# Label the post-processed  probability cube
s2_cube_label <- sits_label_classification(
  cube = s2_cube_bayes,
  output_dir = "./cubes/02_class/",
  memsize = 16,
  multicores = 4
  )

plot(s2_cube_label)

Map uncertainty

To improve model performance, we estimate class uncertainty and plot these pixel error metrics. Results below reveal highest uncertainty levels in classification of wetland and water areas.

# Calculate the uncertainty cube
s2_cube_uncert <- sits_uncertainty(
  cube = s2_cube_bayes,
  type = "margin",
  output_dir = "./cubes/03_error/",
  memsize = 16,
  multicores = 4
)

plot(s2_cube_uncert)

As expected, the places of highest uncertainty are those covered by surface water or associated with wetlands. These places are likely to be misclassified. For this reason, sits provides sits_uncertainty_sampling(), which takes the uncertainty cube as its input and produces a tibble with locations in WGS84 with high uncertainty (Camara et al., n.d.).

# Find samples with high uncertainty
new_samples <- sits_uncertainty_sampling(
  uncert_cube = s2_cube_uncert,
  n = 20,
  min_uncert = 0.5,
  sampling_window = 10
  )

# View the location of the samples
sits_view(new_samples)

Add training samples

We can then use these points of high-uncertainty as new samples to add to our current training dataset. Once we identify their feature classes and relabel them correctly, we append them to derive an augmented samples_round_2.

# Label the new samples
new_samples$label <- "Wetland"

# Obtain the time series from the regularized cube
new_samples_ts <- sits_get_data(
  cube = s2_reg_cube_ro,
  samples = new_samples
  )

# Add new class to original samples
samples_round_2 <- dplyr::bind_rows(
  samples_4classes_3bands,
  new_samples_ts
  )

# Train a RF model with the new sample set
rfor_model_v2 <- sits_train(
  samples = samples_round_2,
  ml_method = sits_rfor()
  )

# Classify the small area cube
s2_cube_probs_v2 <- sits_classify(
  data = s2_reg_cube_ro,
  ml_model = rfor_model_v2,
  output_dir = "./cubes/02_class/",
  version = "v2",
  memsize = 16,
  multicores = 4
  )

# Post-process the probability cube
s2_cube_bayes_v2 <- sits_smooth(
  cube = s2_cube_probs_v2,
  output_dir = "./cubes/04_smooth/",
  version = "v2",
  memsize = 16,
  multicores = 4
  )

# Label the post-processed  probability cube
s2_cube_label_v2 <- sits_label_classification(
  cube = s2_cube_bayes_v2,
  output_dir = "./cubes/05_tuned/",
  version = "v2",
  memsize = 16,
  multicores = 4
  )

# Plot the second version of the classified cube
plot(s2_cube_label_v2)

Remap uncertainty

# Calculate the uncertainty cube
s2_cube_uncert_v2 <- sits_uncertainty(
  cube = s2_cube_bayes_v2,
  type = "margin",
  output_dir = "./cubes/03_error/",
  version = "v2",
  memsize = 16,
  multicores = 4
)

plot(s2_cube_uncert_v2)

Accuracy assessment

To select a validation subset of the map, sits recommends Cochran’s method for stratified random sampling (Cochran 1977). The method divides the population into homogeneous subgroups, or strata, and then applying random sampling within each stratum. Alternatively, ad-hoc parameterization is suggested as follows.

ro_sampling_design <- sits_sampling_design(
  cube = s2_cube_label_v2,
  expected_ua = c(
    "Burned_Area"       = 0.75,
    "Cleared_Area"      = 0.70,
    "Forest"            = 0.75,
    "Highly_Degraded"   = 0.70,
    "Wetland"           = 0.70
  ),
  alloc_options         = c(120, 100),
  std_err               = 0.01,
  rare_class_prop       = 0.1
)
# show sampling desing
ro_sampling_design

Split train/test data

ro_samples_sf <- sits_stratified_sampling(
  cube                  = s2_cube_label_v2,
  sampling_design       = ro_sampling_design,
  alloc                 = "alloc_120",
  multicores            = 4,
  shp_file              = "./samples/ro_samples.shp"
)

sf::st_write(ro_samples_sf,
  "./samples/ro_samples.csv",
  layer_options = "GEOMETRY=AS_XY",
  append = FALSE # TRUE if editing existing sample
)

Confusion matrix

# Calculate accuracy according to Olofsson's method
area_acc <- sits_accuracy(s2_cube_label_v2,
  validation = ro_samples_sf,
  multicores = 4
)
# Print the area estimated accuracy
area_acc

# Print the confusion matrix
area_acc$error_matrix

Times series visualization

summary(as.data.frame(ro_samples_sf))

Deforestation binary map

Deforestation risk map

2. Workflow in Python -> GEE

# Set your Python ENV
Sys.setenv("RETICULATE_PYTHON" = "/usr/bin/python3")
# Set Google Cloud SDK 
Sys.setenv("EARTHENGINE_GCLOUD" = "~/seamus/google-cloud-sdk/bin/")

library(rgee) 
ee_Authenticate()
ee_install_upgrade()
ee_Initialize()

Housekeeping

# convert markdown to script.R 
knitr::purl("VT0007-deforestation-risk-map.qmd")

# display environment setup
devtools::session_info()

References

Camara, Gilberto, Rolf Simoes, Felipe Souza, Felipe Menino, Charlotte Pelletier, Pedro R. Andrade, Karine Ferreira, and Gilberto Queiroz. n.d. Uncertainty and Active Learning | Sits: Satellite Image Time Series Analysis on Earth Observation Data Cubes. https://e-sensing.github.io/sitsbook/uncertainty-and-active-learning.html#uncertainty-and-active-learning.

Cochran, William G. 1977. “Sampling Techniques.” Johan Wiley & Sons Inc.

Stanimirova, Radost, Katelyn Tarrio, Konrad Turlej, Kristina McAvoy, Sophia Stonebrook, Kai-Ting Hu, Paulo Arévalo, et al. 2023. “A Global Land Cover Training Dataset from 1984 to 2020.” Scientific Data 10 (1): 879. https://doi.org/10.1038/s41597-023-02798-5.

Verra. 2021. “VT0007 Unplanned Deforestation Allocation Tool.”

Woodcock, Curtis, Pontus Olofsson, Thomas Loveland, Chris Barber, and Zhe Zhu. n.d. “Global Land Cover Estimation (GLanCE) Product User Guide Version 1.0 August 2022.”